This work proposes a framework that exposes backbone depth V, action expert depth A, and denoising steps $D$ as three jointly configurable compute axes in a VLA, and introduces a KV Cache synthesis mechanism that manages the missing keys and values of the skipped backbone layers, allowing the action expert to exit deeper than the backbone.
Abstract
Flow-matching Vision-Language-Action (VLA) models have emerged as a potential solution for generalist robot control, designed by combining a pretrained Vision-Language Model (VLM) backbone with an action expert that generates continuous robot actions. While these models exhibit impressive capabilities, due to their very high number of parameters, their computational requirements are often prohibitive for robotics control. To mitigate these inefficiencies, existing methods predominantly skip VLM backbone layers with early exits or reduce denoising steps, while leaving action expert depth untouched. We propose a framework that exposes backbone depth $V$, action expert depth $A$, and denoising steps $D$ as three jointly configurable compute axes in a VLA. Starting from a pretrained VLA, we attach lightweight Exit Transformers (ET) at intermediate depths in both the backbone and the action expert, trained to distil the last layer of the policy into each exit. Furthermore, we introduce a KV Cache synthesis mechanism that manages the missing keys and values of the skipped backbone layers, allowing the action expert to exit deeper than the backbone. Finally, we show that the optimal compute budget is task-dependent, with different tasks benefiting from different axes and depths. Notably, our method does not require training the original policy from scratch, and for each exit, it increases the number of parameters by only $2.1\%$ for SmolVLA and $4.1\%$ for $\pi_{0.5}$. We validate our approach across two flow-matching VLAs (SmolVLA, $\pi_{0.5}$) and two benchmarks (LIBERO, Meta-World), revealing complementary effects: $V$ and $A$ respectively reduce FLOPs and latency, while $D$ improves both. Our joint configurations $(V,A,D)$ reduce latency by $79.2\%$ and computation (FLOPs) by $31.8\%$, while improving mean success rate by $5.6\%$.
IMLE-VLA is introduced, which replaces the iterative action head with a single-step conditional generator trained via conditional Implicit Maximum Likelihood Estimation (cIMLE), and promotes multimodal action coverage, avoiding the mode collapse of naive regression heads while eliminating multi-step sampling entirely.
Kian Hosseinkhani, Qin-He Peng, George Shramko et al.· arXiv.org· 2 citations
Robion is designed, the first VLA serving and management system for multi-robot, multi-model requests on multi-GPU edge servers that meets SLOs and enables flexible model placements on multi-GPU servers, and integrates an intelligent traffic controller.
Dionysios Adamopoulos, Nattapol Chanpaisit, Basel Fakhri et al.· 0 citations
The results suggest that dense to MoE adaptation with dynamic expert deactivation is a practical direction for reducing active VLA model size without severe performance loss.
Results show that endpoint correction improves the quality-latency trade-off of frozen flow-matching policies, and supports correction as an effective alternative to additional integration.
Zhi-Peng Tang, Xin-Da Chen, Wei-Ning Rao et al.· 2 citations
This paper proposes TaskAnchor, a lightweight adapter that grounds task state by injecting execution context into the VLA's native input space, and real-robot experiments further validate reliable multi-stage execution.
Heng-Yan Liu, Wen-Lve Zhou, Bo Yue et al.· 0 citations
Flow-based vision-language-action (VLA) models are highly effective for generalist robot manipulation, yet their reliance on computationally expensive VLM encoding and multi-step iterative action generation imposes a significant latency bottleneck. The resulting inference latency makes it difficult for robots to respon...
Yu-Han Chen, Ke Yu, Peng-Fei Liu et al.· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 9, 2026